Mario Alka
@marioalka · Administrator · Member since 07/2026
Ich bin Unternehmer, Softwareentwickler und KI-Enthusiast. Seit vielen Jahren entwickle ich Unternehmenssoftware und beschäftige mich inzwischen fast täglich mit lokalen LLMs, KI-Agenten und leistungsfähiger KI-Hardware.
Mit LLM-Benchmark.de möchte ich eine Plattform schaffen, auf der Modelle, GPUs und Agenten objektiv und reproduzierbar miteinander verglichen werden.
4.603benchmarks contributed
4.552Performance runs
51Harness runs
88Models tested
20GPUs used
2.491Best throughput · tok/s
100,0%Best success rate
Hardware used
NVIDIA RTX A60001019× · 1.568 tok/sAMD Radeon AI PRO R9700975× · 667 tok/sNVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition560× · 2.144 tok/sAMD Radeon RX 7900 XTX338× · 718 tok/sNVIDIA RTX PRO 6000 Blackwell Workstation Edition244× · 2.414 tok/sAMD Radeon PRO W7900 Dual Slot241× · 344 tok/sNVIDIA GeForce RTX 5070 Ti222× · 1.744 tok/sNVIDIA GeForce RTX 3090 Ti222× · 1.493 tok/sNVIDIA GeForce RTX 5090218× · 2.491 tok/sAMD Radeon PRO W7800 48GB149× · 703 tok/sNVIDIA GB1087× · 903 tok/sNVIDIA GeForce RTX 206077× · 260 tok/s
Models tested
Qwen3.6-35B-A3B160×gpt-oss-20b151×Qwen3-30B-A3B-Instruct-2507130×Laguna-XS-2.1127×Qwen3-Coder-30B-A3B-Instruct126×Ministral-3-14B-Reasoning-2512123×Qwen3-30B-A3B-Thinking-2507115×Qwen3.5-35B-A3B111×Meta-Llama-3.1-8B-Instruct111×Ornith-1.0-9B109×Devstral-Small-2-24B-Instruct-2512105×Qwen2.5-Coder-32B-Instruct104×Qwen3-VL-30B-A3B-Instruct102×Magistral-Small-2509101×
Top performance runs
| Model | Hardware | tok/s | TTFT | Date |
|---|---|---|---|---|
| gemma-4-E2B-it 10× | NVIDIA GeForce RTX 5090 llama.cpp | 2.491 | 4.382 ms | Jul 28, 2026 |
| gemma-4-E2B-it 10× | NVIDIA RTX PRO 6000 Blackwell Workstation Edition llama.cpp | 2.414 | 4.107 ms | Jul 27, 2026 |
| Nemotron-3-Nano-4B 10× | NVIDIA RTX PRO 6000 Blackwell Workstation Edition llama.cpp | 2.183 | 5.598 ms | Jul 27, 2026 |
| gemma-4-E2B-it 10× | NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition llama.cpp | 2.144 | 4.819 ms | Jul 28, 2026 |
| gpt-oss-20b 10× | NVIDIA RTX PRO 6000 Blackwell Workstation Edition llama.cpp | 1.995 | 5.309 ms | Jul 29, 2026 |
| gpt-oss-20b 10× | NVIDIA RTX PRO 6000 Blackwell Workstation Edition llama.cpp | 1.994 | 5.323 ms | Jul 28, 2026 |
| Nemotron-3-Nano-4B 10× | NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition llama.cpp | 1.938 | 6.288 ms | Jul 28, 2026 |
| Nemotron-3-Nano-4B 10× | NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition llama.cpp | 1.912 | 6.245 ms | Jul 28, 2026 |
| Nemotron-3-Nano-4B 10× | NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition llama.cpp | 1.897 | 6.227 ms | Jul 27, 2026 |
| gpt-oss-20b 10× | NVIDIA GeForce RTX 5090 llama.cpp | 1.891 | 5.489 ms | Jul 28, 2026 |
Harness results
| Model | Hardware | Success rate | Points | Date |
|---|---|---|---|---|
| Qwen3.6-35B-A3B-AWQ | NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition | 100,0% | 200/200 | Sep 29, 2026 |
| Jev-1.13.0 | 100,0% | 200/200 | Sep 29, 2026 | |
| Qwen3.6-35B-A3B-AWQ | NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition | 100,0% | 200/200 | Sep 29, 2026 |
| Qwen3.8-27B-Uncensored | NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition | 100,0% | 200/200 | Sep 29, 2026 |
| Open-Jev-27B-v1.1 | NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition | 100,0% | 200/200 | Sep 29, 2026 |
| Kolibri-1 | NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition | 100,0% | 100/100 | Oct 8, 2026 |
| Kolibri-1 | NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition | 98,6% | 207/210 | Oct 8, 2026 |
| Qwen3.6-35B-A3B-AWQ | NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition | 96,5% | 193/200 | Sep 29, 2026 |
| Kolibri-1 | NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition | 93,2% | 289/310 | Oct 8, 2026 |
| Kolibri-1 | NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition | 91,3% | 283/310 | Oct 8, 2026 |